App onboarding clip from screenshots: Wan 3.0 takes 10 references

Use Wan 3.0 and send up to 10 of your app screens as reference images. A 10-second 9:16 clip is $0.63 at 480p, $1.25 at 720p and $2.50 at 1080p on Sume. Kling 3 is not an option for this on Sume: it takes a first and last frame but no reference images.

What the model does with screens

Alibaba's Wan 3.0 launch post, read 2026-10-08, lists text, image, audio, video and document inputs, and describes a one-file-or-link-per-request document input of up to 100 MB and 50 pages. Sume does not expose the file or web link input; its Wan 3.0 entry lists text, image, reference images, reference video and reference audio instead.

Treat the output as a promotional animation around the screens, not a faithful recording. Alibaba says on-screen text accuracy is still improving, so a generated phone UI will drift. For an onboarding clip, that means the real screen recording belongs in the video, and the generated part is the context around it: a hand, a desk, a commute.

Prices and limits

Sume bills the provider list price times 1.25, rounded up to the cent per job.

How to run it cheaply

Reference-heavy requests are best drafted small.

• Pin resolution explicitly. Omitted, Wan 3.0 is priced and rendered at 1080p.
• Send 3 to 4 screens that show the key steps, not all 10.
• Reference videos are capped at 5 files and 15 seconds in total, and need at least 16 fps. Reference audio is capped the same way.
• Start at 480p, $0.63, and look at pacing before the 720p render.

Submit it and read the result

Sume reserves the billable amount from your workspace USD balance when the job is submitted, then returns a polling_url. Add up to 10 reference images through input_references entries of type image_url. Poll the job until status is completed or failed; a completed job lists its files in unsigned_urls.

import os
import time
import requests

H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
body = {
    "model": "wan-3.0",
    "prompt": "A slow push-in on the scene, soft natural light",
    "duration": 10,
    "resolution": "480p",
    "aspect_ratio": "9:16",
}
r = requests.post("https://api.sume.com/v1/videos", headers=H, json=body, timeout=60)
r.raise_for_status()
url = r.json()["polling_url"]
while True:
    job = requests.get(url, headers=H, timeout=30).json()
    if job["status"] in ("completed", "failed"):
        break
    time.sleep(10)
print(job["status"], job.get("unsigned_urls"))

What not to expect

I did not test how faithfully Wan 3.0 reproduces interface layouts, and the vendor pages I read make no claim about it. Sume also does not expose Wan's thinking toggle. If your store listing needs pixel-exact UI, use a screen recording and use the model only for lifestyle shots.